tarekmasryo/llm-system-ops-production-telemetry-sft-data

🤖📈 LLM System Ops Telemetry (Synthetic)

A synthetic, production-style, multi-table LLM telemetry dataset designed for LLMOps analytics and decision-grade experiments.

It supports monitoring cost, latency, tokens, failures, safety flags, tool usage, and user feedback at the interaction level, with rollups at the session and user levels — plus an SFT table aligned 1:1 with interactions and a prompt/config dimension.

• Synthetic data (safe for teaching, prototyping, and portfolio notebooks).
• Not real user data.
• cost_usd and token fields are synthetic estimates (not billing truth).

---

📌 Dataset summary

• Time coverage (UTC): 2025-02-01 → 2025-04-30
• Core grain: 1 row = 1 LLM interaction (request → response)
• Tables (configs): interactions, sessions_summary, users_summary, prompts_lookup, sft
• Split: split is deterministic and group-safe, derived from session_id hashing (prevents session leakage)

---

🧾 What’s inside

Files

• data/llm_system_interactions.csv — interaction-level telemetry (core fact table)
• data/llm_system_sessions_summary.csv — session-level aggregates
• data/llm_system_users_summary.csv — user-level aggregates
• data/llm_system_prompts_lookup.csv — prompt/config dimension table keyed by prompt_id
• data/llm_system_instruction_tuning_samples.csv — instruction-tuning samples aligned 1:1 with interactions
• docs/data_dictionary.md — schema, types, formulas, allowed values, and leakage notes

Repository layout

.
├─ data/
│  ├─ llm_system_interactions.csv
│  ├─ llm_system_sessions_summary.csv
│  ├─ llm_system_users_summary.csv
│  ├─ llm_system_prompts_lookup.csv
│  └─ llm_system_instruction_tuning_samples.csv
└─ docs/
   └─ data_dictionary.md

---

🔗 Relational structure (keys & joins)

llm_system_users_summary (user_id)
  └── llm_system_sessions_summary (session_id, user_id)
        └── llm_system_interactions (interaction_id, session_id, user_id, prompt_id)
              └── llm_system_instruction_tuning_samples (sample_id, interaction_id, prompt_id)

llm_system_prompts_lookup (prompt_id)  ──<  llm_system_interactions.prompt_id

Primary keys

• users_summary: user_id
• sessions_summary: session_id
• interactions: interaction_id
• prompts_lookup: prompt_id
• sft: sample_id

Foreign keys

• sessions_summary.user_id → users_summary.user_id
• interactions.session_id → sessions_summary.session_id
• interactions.user_id → users_summary.user_id
• interactions.prompt_id → prompts_lookup.prompt_id
• sft.interaction_id → interactions.interaction_id (1:1)
• sft.prompt_id → prompts_lookup.prompt_id (many-to-one)

---

🧠 Important semantics

Text fields (preview vs full text)

• interactions contains only a preview:

  • response_text_snippet
  • response_text_snippet_len_chars
  • response_text_snippet_is_truncated

• Full text lives in sft:

  • raw: raw_request_text, raw_response_text
  • SFT-ready: sft_user_prompt, sft_assistant_response

Group-safe split

Use split for train/val/test evaluation. Avoid random row-level splits, or you may leak session patterns across splits.

Leakage note (modeling)

Some observability fields are redundant by design. Example: when predicting is_failure, drop failure_type and failure-related flags (and any post-resolution fields).

---

🛠️ How to use

Load with Datasets

from datasets import load_dataset, get_dataset_config_names

repo_id = "tarekmasryo/llm-system-ops-production-telemetry-sft-data"

print("Configs:", get_dataset_config_names(repo_id))

# default config (interactions)
interactions = load_dataset(repo_id)["train"].to_pandas()

# specific tables
sessions = load_dataset(repo_id, "sessions_summary")["train"].to_pandas()
users = load_dataset(repo_id, "users_summary")["train"].to_pandas()
prompts = load_dataset(repo_id, "prompts_lookup")["train"].to_pandas()
sft = load_dataset(repo_id, "sft")["train"].to_pandas()

print(interactions.shape, sessions.shape, users.shape, prompts.shape, sft.shape)

Clone the dataset repository (local)

git clone https://huggingface.co/datasets/tarekmasryo/llm-system-ops-production-telemetry-sft-data
cd llm-system-ops-production-telemetry-sft-data

Local load (after cloning)

import pandas as pd

interactions = pd.read_csv("data/llm_system_interactions.csv")
sessions = pd.read_csv("data/llm_system_sessions_summary.csv")
users = pd.read_csv("data/llm_system_users_summary.csv")
prompts = pd.read_csv("data/llm_system_prompts_lookup.csv")
sft = pd.read_csv("data/llm_system_instruction_tuning_samples.csv")

Example join

df = (
    interactions
    .merge(sessions[["session_id", "failure_rate", "total_requests"]], on="session_id", how="left")
    .merge(prompts[["prompt_id", "instruction_template", "use_case"]], on="prompt_id", how="left")
)

print(df.shape)
print(df["split"].value_counts(dropna=False))

---

🎯 Suggested use cases

• Monitoring & observability: trends for cost/latency/tokens/failure rate; peak-hour slices; alert rules
• Failure RCA: tool_error vs hallucination vs safety_block; retries and escalation effectiveness
• Cost–quality trade-offs: model comparisons under quality × cost × latency; rollout gating policies
• Tool usage analytics: impact of tools on quality/failure/latency; tool-error patterns
• Feedback analysis: CSAT by use_case/tier/region; response rates; negative feedback investigation
• SFT workflows: align telemetry with the 1:1 SFT table for instruction-tuning experiments

---

🗒️ Notes

• This dataset is fully synthetic and generated for benchmarking/education/EDA workflows.
• Token and cost fields are synthetic estimates (not billing truth).
• Refer to docs/data_dictionary.md for detailed schema, formulas, allowed values, and leakage guidance.

---

🧾 Citation

Tarek Masryo. (2025). LLM System Ops Telemetry (Synthetic).

---

📜 License

CC BY 4.0 (Attribution) — free to share and adapt with attribution.